idkmesh

IDKMesh Randomness / Bio-Inspired Experiment Status

Updated: 2026-08-29 Program rule: Randomness controls exploration, not acceptance.

This page is a compact status map for the executable research program derived from RANDOMNESS_AND_BIOINSPIRED_ALGORITHMS.md and docs/research/RANDOMNESS_ROADMAP.md.

The purpose is to keep contributors from having to reconstruct the experiment graph from many issues and pull requests.

Governing architecture

safe / deterministic envelope
          +
stochastic exploration and diversity
          +
independent verification
          +
evidence-backed selection
          +
persistent memory
          ->
adaptive collective search

Randomness may decide:

what to explore;
who attempts work;
which local candidates receive resources;
which independent verifier is assigned;
how non-critical information propagates;
how experimental policy variants are generated.

Randomness must not substitute for:

authorization;
provenance;
safety gates;
independent verification;
canonical acceptance.

Current experiment graph

#29 randomness-lab foundation                    COMPLETE
      |
      +-- #30 R1 swarm diversity                MECHANISM COMPLETE / REAL EVIDENCE OPEN
      |      |
      |      +-- #70 real R1 coding corpus      OPEN
      |
      +-- #31 R2 randomized scheduling          COMPLETE
      |      |
      |      +-- #84 stronger R2 evidence       OPEN
      |
      +-- #32 R3 evolutionary orchestration     SYNTHETIC COMPLETE
      |      |
      |      +-- #96 real-task R3 phase         OPEN
      |
      +-- #97 R4 verified stigmergic routing    SYNTHETIC COMPLETE

R0 — randomness-lab foundation

Issue: #29 — completed
Core PR: #39

Implemented reusable dependency-free experimental infrastructure:

deterministic seeded simulation;
pluggable outcome environments;
greedy baseline;
epsilon-greedy;
softmax/Boltzmann selection;
UCB;
Thompson sampling;
power-of-d least-loaded primitive;
repeated trials;
raw machine-readable metrics;
descriptive uncertainty summaries;
Python 3.11–3.13 CI.

This remains the common substrate for the randomness experiments.

R1 — diversity, error correlation, and verification

Issue: #30 — mechanism implemented; open pending real corpus
Implementation PRs: #54, #58, #67
Real evidence: #70

Question

When does stochastic/structural diversity improve verified coding outcomes compared with repeated copies of the same worker?

Implemented

Six-condition synthetic harness

one deterministic worker;
identical replication;
seed-only variation;
structural diversity;
Thompson-selected workers;
structural diversity + randomized verifier assignment.

Help/hurt map

The parameter sweep explicitly reports:

helps
hurts
uncertain

instead of tuning only for positive diversity results.

Variables include:

worker error correlation;
verifier correlation;
swarm size;
structural-worker quality penalty.

Real-result replay adapter

randomness_lab.r1_replay consumes the existing Phase 0:

ResultManifest v0.1;
independently verified VerificationResult v0.1.

Worker self-report alone never establishes truth.

The replay compares equal candidate budgets and can measure actual structural-signature failure correlation.

Real-corpus readiness gate

randomness_lab.r1_readiness now fails closed before a BenchmarkCohort is interpreted as real R1 evidence. It checks held-out/frozen status, the prospective 20-work-unit target, exact per-signature candidate budgets, conclusive independent verification, replay/cohort signature agreement, independent tests, retained seeded negatives, complete cost measurements, and non-selecting authority. The committed current-state audit is blocked at zero eligible work units; it is explicitly contract-state evidence, not a coding outcome.

Evidence boundary

The repository does not yet contain a sufficient real multi-worker corpus.

#70 therefore blocks a real empirical answer. It requires held-out coding work, multiple baseline replicas, multiple structural signatures, independent verification, retained negative results, and public reproducible evidence.

R2 — randomized scheduling under churn

Issue: #31 — completed
Implementation PRs: #77, #80
Reference evidence PR: #83
Stronger follow-up: #84 — synthetic Phase A/B/C completed

Question

Can bounded local random choices produce useful scheduling quality without maintaining globally current state?

Policies

one random choice;
power-of-two;
power-of-three;
capability-aware power-of-two;
global least-loaded high-information oracle.

Model

The simulator has replayable task/outage traces with:

heterogeneous capacity;
heterogeneous capabilities;
bursty arrivals;
stale availability;
stale load observations;
worker outages/churn;
capability mismatch;
locality mismatch;
requeue/recovery;
uncheckpointed lost work.

The oracle is charged for full capability-pool metadata scans.

Published reference scale

results/experiments/r2/reference-scale-seed42.json
results/experiments/r2/reference-scale-seed42.md

Scales:

1
10
100
1,000
10,000
100,000 workers

Regimes:

fresh
moderate
stale

First single-seed observations

These are reference observations, not universal conclusions.

All published cells completed all tasks within the configured drain horizon.
Capability-aware power-of-two generally stayed close to oracle p95 response while using roughly 2 metadata probes per routing attempt.
The oracle metadata scan grew from roughly 5 probes/attempt at 10 workers to roughly 4.4k probes/attempt at 10k workers.
At 100k workers the full oracle was intentionally skipped; bounded local policies remained executable.
Capability-oblivious policies produced many more failed assignments, especially at large scale.
Some small moderate/stale cells triggered the explicit loses_badly diagnostic, proving the harness can expose local-routing failure regimes.

Stronger evidence completed

#84 now retains:

a five-seed full scale ladder;
factor-isolated capability-rarity sweeps;
independent availability/load staleness sweeps;
matched independent and regional failure shapes;
workload saturation separate from fleet-size scaling;
deterministic communication, directory, and state-cost proxies;
an explicitly host-specific scheduler CPU/peak-memory profile.

The result is a regime map, not a universal power-of-two claim. Real network and fleet measurements remain outside this synthetic evidence boundary.

R3 — evolutionary orchestration

Issue: #32 — synthetic mechanism experiment completed
Implementation PR: #93
Frozen evidence PR: #95
Real-task phase: #96

Question

Can evolutionary search discover useful orchestration policies without using held-out results during selection, collapsing into one monoculture, or autonomously promoting itself?

Genome

worker count;
structural diversity mix;
decomposition depth;
replication factor;
verifier depth;
exploration temperature;
timeout budget;
escalation threshold.

Search

changing training distributions;
Pareto multi-objective selection;
novelty-aware tie-breaking;
bounded diversity archive;
mutation;
crossover;
random-immigrant exploration floor;
fixed baseline;
fixed final training reference;
pre-heldout champion selection;
held-out evaluation only after search ends.

Objectives

Maximize:

verified success;
worst-family success.

Minimize:

security failures;
regressions;
error correlation;
compute;
latency;
human attention;
complexity.

Raw output volume/activity is not a fitness objective.

Frozen seed-42 evidence

results/experiments/r3/reference-seed42.json
results/experiments/r3/reference-seed42.md

Split digest:

sha256:261e1edd128ee0492fd5b740a1576a0eeb1c5ef4cfb6ccbdf90989fac3f610f5

Pre-heldout champion:

g-7d18f6c8917d

Held-out reference result:

Metric	Fixed baseline	Pre-heldout champion
Verified success	0.2917	0.6417
Security failure	0.0583	0.0208
Regression	0.0792	0.0542
Compute/task	5.6048	8.1963
Latency/task	8.3435	10.6808
Human attention/task	0.2525	0.0833

Important mixed evidence:

47 final train-Pareto genomes reached held-out evaluation;
12 triggered the configured overfit flag;
champion train→heldout success gap = 0.0733;
champion held-out success delta vs baseline = +0.3500;
champion security-failure delta vs baseline = -0.0375;
champion costs more compute and latency;
the strongest recommendation is only consider_for_separate human-reviewed experiment;
autonomous promotion is always false.

The synthetic held-out split is now burned for confirmatory tuning.

#96 requires a new real-task train/held-out split and actual Verified Swarm Runner controls/outcomes.

R4 — verified stigmergic routing

Issue: #97 — synthetic mechanism and frozen reference complete

Question

Can ant-colony-like traces learn task→worker affinities from verified outcomes, adapt when capabilities change, and still give newcomers enough opportunities to demonstrate value?

Candidate rule:

tau[a,i](t+1) = (1-rho) * tau[a,i](t) + verified_deposit[a,i](t)

Critical constraints:

activity does not produce pheromone;
verified durable outcome does;
old evidence evaporates;
newcomer/exploration floor remains non-zero;
canonical acceptance remains independent from routing weight;
no direct conversion of pheromone into governance power.

Implemented baselines:

uniform random;
greedy empirical success;
Thompson sampling;
stigmergy without evaporation;
stigmergy with evaporation;
stigmergy + evaporation + exploration/newcomer floor.

Implemented stressors:

specialization;
capability shift;
newcomer arrival;
misleading early success;
churn;
lock-in.

Frozen reference evidence

results/experiments/r4/reference-default.json — full 800-step routing traces, metrics, and pheromone snapshots;
results/experiments/r4/reference-lockin.json — full 500-step adversarial lock-in traces, metrics, and snapshots;
results/experiments/r4/reference-summary.md — commands, digests, comparison, limitations, and interpretation.

The adaptive stigmergic policy avoided the catastrophic permanent-pheromone lock-in and explored both strong and weak newcomers. Thompson sampling still slightly led realized success in the default trace and decisively won the lock-in trap. The biological analogy therefore does not establish superiority.

Every stigmergic reference run records exactly zero pheromone increase from unverified activity. A deterministic regression test verifies both artifact hashes, preserves cross-runtime replay invariants and the negative regime, and requires byte identity on the recorded Python 3.12 runtime family.

Next-order biological tracks

These remain hypotheses rather than implementation commitments.

Honeybee-style cross-inhibition

Use competing evidence accumulation plus explicit counter-evidence to test whether architecture/proposal decisions avoid deadlock and premature consensus better than majority voting.

Quorum sensing

Test local thresholds for activating expensive collective actions, for example:

starting large integration work;
spawning an autonomous cell;
escalating verification;
promoting an experimental protocol to a broader trial.

A quorum is a coordination threshold, not proof of correctness.

Physarum-style adaptive topology

Strengthen communication/task routes carrying verified useful flow and decay weak routes while preserving redundancy.

Candidate use: evolve node -> cell -> region -> federation connectivity under cost/latency/fault-tolerance objectives.

Artificial immune mechanisms

Use distributed detector memory as an experimental security/anomaly layer, benchmarked against conventional anomaly-detection baselines rather than assuming biological inspiration is superior.

What we have learned so far

1. Diversity is not agent count

The useful variable is closer to:

independent useful attempts
- correlated failure
- compute cost
- human attention

than raw worker count.

2. Information has a scalability cost

R2 shows why scheduler quality must be reported together with the amount of global state consulted.

A bounded capability index + tiny candidate sample is currently a more interesting IDKMesh hypothesis than either uniform random assignment or a globally current full scan.

3. Search needs a test it cannot rewrite

R3 makes the held-out boundary operational. Evolution can search aggressively, but its candidate must be chosen before confirmatory evidence is opened.

4. Negative results are part of the product

The experiment APIs explicitly preserve:

harmful diversity cells;
local scheduling regimes that lose to the oracle;
overfit evolved genomes;
failed/security/regression evidence.

A research system that deletes those outcomes cannot reliably self-improve.

5. Memory needs forgetting

The next R4 question makes this explicit. Permanent historical advantage creates lock-in; zero memory wastes evidence. Controlled evaporation/decay is likely to be a core parameter across routing, reputation, topology, and community systems.

6. Community algorithms need stronger abuse resistance than routing algorithms

A pheromone value that routes a synthetic task is low risk. A score that influences contributor authority or resource ownership is much more dangerous.

Do not reuse experimental routing metrics as governance power without separate work on:

Sybil resistance;
collusion;
newcomer fairness;
capture;
metric gaming;
appeals;
transparent decay;
human governance.

Immediate contributor path

If you want to contribute to this program now:

Real coding evidence: #70 (R1).
R2 real-fleet validation: open a bounded follow-up from completed #84 evidence.
Real evolutionary orchestration evidence: #96 (R3).
New bio-inspired algorithm implementation: #97 (R4 stigmergy).

The experiments are intentionally separable. A contributor should not need to understand all of IDKMesh before improving one policy, environment, metric, or evidence pipeline.

This site is open source. Improve this page.